Florin Iluta

Systems. AI software.

[Florin Iluta @FlorinIluta]

The better coding agents get, the less time I spend worrying about getting code written.

More of the work moves to scope, taste, checks and handoffs.

[Florin Iluta @FlorinIluta]

@NotionHQ Same skill and same approved files across agents also makes comparisons a lot fairer. You're removing one big variable and seeing more clearly what each agent actually does with the same input.

[Florin Iluta @FlorinIluta]

@sircalebhammer I've found the biggest savings happen before execution. Good research, a tight scope and a clear plan let you build a lot without throwing tokens at every problem.

[Florin Iluta @FlorinIluta]

@sairahul1 This matches what I've seen: once planning, critique and execution are split across models, the final review matters more, not less. One early mistake can survive every handoff and still look convincing at the end.

[Florin Iluta @FlorinIluta]

@apify @X @eptwts A 14K-tweet archive is basically a dataset of your own thinking. At that size, an agent can start finding patterns and ideas you probably forgot you ever wrote.

[Florin Iluta @FlorinIluta]

@aacle_ This carries well beyond security. A broad pass can map the problem, but once you know where to look, giving the agent one thing to chase usually gets you much deeper.

[Florin Iluta @FlorinIluta]

@MengTo This solves one of the harder parts of agent-built UI: starting with good taste. Starting from something this polished makes a huge difference.

[Florin Iluta @FlorinIluta]

@testingcatalog Custom agents on mobile is the bigger one for me. If the same agents carry over from web, they become a lot more useful between desktop sessions.

[Florin Iluta @FlorinIluta]

@Dimillian How far can you take the loop without opening Xcode? If server-sim is good enough for most UI iteration, that's a pretty big difference.

[Florin Iluta @FlorinIluta]

@jetbrains I use different models depending on the task. The less my workflow depends on one specific tool, the easier it is to switch.

[Florin Iluta @FlorinIluta]

@orhundev Using Claude's own transcript for this is nice. You get the full replay without adding another layer to the agent loop.

[Florin Iluta @FlorinIluta]

@hey_mujeebahmed Pre-wiring the infrastructure also removes a lot of decisions from the agent loop. Less room to invent a different architecture every time.

[Florin Iluta @FlorinIluta]

@askOkara Live GA data and Competitor Pulse together make sense. Having your own numbers and competitor changes in one place saves a lot of tab hopping.

[Florin Iluta @FlorinIluta]

@JFPuget Built for GPU kernels, then it handles all 183 of these with almost no setup. That's a big jump for the same system.

[Florin Iluta @FlorinIluta]

@DanKornas The raw vs consolidated split is smart. Not everything remembered during a session deserves to become permanent.

[Florin Iluta @FlorinIluta]

@kawai_design そうですね。実際のUI���そのまま教材になるのがいいですね。

[Florin Iluta @FlorinIluta]

@mave99a Five screenshot rounds is probably the signal to change the approach, not add a sixth.

[Florin Iluta @FlorinIluta]

@galfrevn Record the real session once, then make the demo look good afterward. No reason to rerun the work just for the presentation.

[Florin Iluta @FlorinIluta]

@nickbaumann_ The heartbeat opens up more than the DJ use case. Give an agent one small thing to keep an eye on and it stops being a one-shot task.

[Florin Iluta @FlorinIluta]

@kawai_design Docs are usually the first thing to drift. Generating the guide from the real UI is a pretty good way to keep them in sync.